公开资料显示，壁仞科技为国产AI芯片核心企业。公司在2025年实现了BR166芯片的量产。壁砺166L及 壁砺166M已于2025年8月开始量产，壁砺166C已于2025年12月开始量产。下一代产品BR20X预计于 2026年量产出货。 消息面上，2026年春节前后，中国国产AI大模型密集发布。壁仞科技旗舰产品壁砺166系列已完成对智 谱GLM-5、MiniMax M2.5、DeepSeek、千问Qwen3.5等头部模型的高效兼容，凭借高算力、高通用性与 全栈自研软件栈，成为支撑百模大战的核心算力底座。 壁仞科技(06082)再涨近5%，截至发稿，涨4.05%，报40.58港元，成交额3.52亿港元。 ...

图片来自Taalas官网 打着"颠覆英伟达"的旗号的公司，总是会接二连三涌现。 Taalas称，公司通过结构化ASIC技术将芯片定制周期缩短至两个月，已累计融资2.19亿美元。24名敬业 的员工的努力，投入3000万美元，打造出了这款拥有"极致的专业化、速度和能源效率"的产品。 Taalas创始人兼CEO是曾任AMD架构师的业界传奇人物柳比沙·巴伊奇（Ljubiša Bajić）。 在公司官网的介绍文章中，巴伊奇称，这款芯片选择了Meta公司2024年7月推出的开源大模型Llama 3.1 8B作为运行平台，峰值推理速度接近17000 tokens/秒，比目前市场中最先进的技术快近10倍，构建成本 降低到原来的1/20，功耗降低至原来的1/10。 截图来自社交平台X 巴伊奇给出了一组测试数据，Taalas自己在Llama 3.1 8B上测试了英伟达的主力产品H200和B200，结果 为230 tokens/秒和353 tokens/秒，而Taalas的HC1性能是它们的48倍。 最近，一家来自加拿大多伦多的芯片"小厂"Taalas引起了AI圈关注。有声音认为，它很可能撬动英伟达 主宰了多年的AI芯片市场。 当 ...

直接把大模型装在芯片面上。。。 这样什么内存带宽限制之类都不存在了。。。 而且性能非常夸张： 在这块专用芯片上跑Llama 3.1，速度能达到17000 token/秒。。。 英伟达最强的H200大概230 token/秒。。。 英伟达超级大芯片B200也就2000 token/秒。。。 来源：贩财局 这个春节，有个叫Taalas的芯片新公司横空出世，它干了一件让英伟达震惊的事情： 这个差距不是一点半点，是直接降维打击。。 想体验可以到这里：chatjimmy.ai 我试了一下，感受到快、非常快、恐怖的快。。。不亚于第一次使用ChatGPT的震撼 按照现在发展速度，很快实现22,000 tokens/s，意味着AI的反应速度已经超过了人类神经的传导速度。 我总结它的优点： 每次升级都得换硬件。。。。 打个比方，传统GPU像是任天堂Switch游戏机，卡槽一插，想玩《塞尔达》玩《塞尔达》，想玩《马里 奥》玩《马里奥》，换游戏只需要换卡带。 Taalas相当于直接把《塞尔达》这部游戏烧死在主板上。想玩《马里奥》？没法下载，也没法插卡，只 能把整台机器扔掉，买一台出厂就烧好《马里奥》的新机器。 不管怎么样，Taal ...

据最新环评资料显示，「台南科学工业园区特定区（开发区块A）新建半导体厂计划」环境影响说明书 公开会议预计在2月24日举行。这块被外界视为加码兴建一座2nm晶圆厂的计划基地位于台南市安定 区，台南科学工业园区特定区之开发区块A范围，南侧临接「树谷园区」，东侧临接南科「台南园 区」。 台积电最新公布的新建厂计划环说书，正式公布计划基地面积约15.46公顷，同时计划期程也正式拍 板。据计划指出，预计2026年启动建筑规划、执照申请及现地开工等作业，2028年办理竣工及申领使 照，预定办理情形将依先期评估及工程作业等实际完成时程调整。 为应对AI芯片订单需求热，台积电规划继续加码投资2nm厂的脚步不停歇。据环评开发进度显示，台积 电去年提出在南科特定区A区区块新建半导体厂环评计划，基地面积约15.46公顷，规划2026年启动建筑 规划、执照申请及现地开工等作业，预计2028年完工启用。 英伟达（NVIDIA）CEO黄仁勋1月表示，英伟达今年需求量很大，更预期未来十年台积电产能可能增 长超过100％。台积电近年扩厂脚步不停歇，去年也提出在南科特定区A区区块新建半导体厂环评计 划，外传是为纾缓2nm订单。 ...

周五（2月20日）， 美国最高法院裁定特朗普的关税政策违法， 令美股低开高走， 三大指数集体收涨。 截至收盘，道琼斯指数涨0.47%，报49625.97点；标普500指数涨0.69%，报6909.51点；纳斯达克综合指数涨0.9%，报22886.07点。 | 名称 | 最新价 | 新大都 | 新城區 | | --- | --- | --- | --- | | 道瑪斯指数 | 49625.97 | +230.81 | 0.47% | | 标普500指数 | 6909.51 | +47.62 | 0.69% | | 纳斯达克综合指数 ♥ 22886.07 | | +203.34 | 0.90% | 当天，美国联邦最高法院大法官以6比3的结果维持下级法院裁决，认定美国总统特朗普援引《国际紧急经济权力法》（IEEPA）实施关税政 策超出了其法定权限。 尽管华尔街先前已普遍预期最高法院会做出不利于特朗普政府的裁决，但仍有一些问题悬而未决，例如已征收关税是否退还及如何退还，这 一问题可能需由下级法院进一步裁定。 FBB Capital Partners高级研究分析师Michael Brenner表示："下级法院将不得不弄 ...

来源：华尔街见闻 英伟达CEO黄仁勋上周亲自出面宴请合作伙伴的工程师团队，这一罕见举动凸显出下一代高带宽内存 HBM4对这家AI芯片巨头的战略重要性。 随着三星电子在HBM4竞赛中率先出货，SK海力士能否按时交付高性能产品，将直接影响英伟达今年 下半年推出的新一代AI加速器Vera Rubin的市场表现。 2月14日晚间，黄仁勋现身英伟达总部附近的一家韩式炸鸡餐厅，花费约两小时为30多名SK海力士和英 伟达工程师逐桌调制烧啤、敬酒致意。 这场临时安排的聚会释放出明确信号：HBM4被英伟达视为Vera Rubin加速器的关键差异化组件。 英伟达对HBM4供应商提出了"运行速度11 Gbps以上"和"带宽3.0 TB/s以上"的规格要求，较竞争对手 AMD的同类要求高出30%以上。 行业分析师认为，黄仁勋的亲自现身提升了SK海力士维持HBM4最大供应商地位的可能性。 据行业消息人士透露，英伟达去年12月初步分配了今年的HBM供应份额，SK海力士获得55%以上，三 星电子占20%多至接近30%，美光科技约20%。虽然三星在技术竞赛中的推进引发了份额变动的可能， 但业内普遍认为SK海力士有望在第一季度完成质量优化后确 ...

英伟达CEO黄仁勋罕见"亲自出马"，在硅谷韩式餐厅宴请30多名工程师，敦促SK海力士按时交付第六 代HBM4。随着三星率先出货搅动战局，这场关乎下一代AI芯片Vera Rubin成败的供应链争夺战已进 入白热化，SK海力士能否守住55%的市场份额，未来数月的执行力将是生死关键。 英伟达CEO黄仁勋上周亲自出面宴请合作伙伴的工程师团队，这一罕见举动凸显出下一代高带宽内存 HBM4对这家AI芯片巨头的战略重要性。随着三星电子在HBM4竞赛中率先出货，SK海力士能否按时 交付高性能产品，将直接影响英伟达今年下半年推出的新一代AI加速器Vera Rubin的市场表现。 2月14日晚间，黄仁勋现身英伟达总部附近的一家韩式炸鸡餐厅，花费约两小时为30多名SK海力士和 英伟达工程师逐桌调制烧啤、敬酒致意。他反复强调"我们是一个团队"和"我为你们骄傲"，并敦促工 程师们"通过不懈挑战和努力交付非凡成果"， 特别提及SK海力士承诺供应的第六代HBM4产品。 这场临时安排的聚会释放出明确信号：HBM4被英伟达视为Vera Rubin加速器的关键差异化组件。英 伟达对HBM4供应商提出了"运行速度11 Gbps以上"和"带宽3.0 ...

Core Viewpoint - The 2026 Haidian New Spring Technology Fair showcases the integration of technology and cultural traditions, featuring over 70 companies and 150 interactive exhibits, emphasizing AI and robotics innovations [1] Group 1: Event Overview - The event is themed "Qiji Chicheng, AI Celebrating the New Year in Haidian" and is held at the Wukesong Wanda Plaza [1] - The fair runs from February 19 to February 23, 2026, and is free to the public [1] Group 2: Exhibits and Features - The fair includes 43 embodied robot exhibits, with 34 being humanoid robots, alongside various AI models, digital humans, VR, and AIGC interactive experiences [1] - It features cutting-edge technology from international events like CES 2026 and the World Robot Conference, highlighting Haidian's innovation and manufacturing capabilities [1] Group 3: Thematic Focus - The event aims to create a cultural symbol of Haidian by combining traditional New Year customs with technology experiences and future lifestyle concepts [1] - Five distinct exhibition areas are designed to enhance consumer engagement across six floors of the Wukesong Wanda Plaza [1]

Core Viewpoint - Nvidia's CEO Jensen Huang announced the unveiling of a "world unprecedented" new chip at the upcoming GTC 2026 conference, which is expected to further solidify the company's leading position in the AI infrastructure sector [1][3]. Group 1: Upcoming GTC 2026 Conference - The GTC 2026 conference will take place on March 15 in San Jose, California, focusing on a new era of AI infrastructure competition [3]. - Huang acknowledged the challenges in developing these new chips, stating that "all technologies are approaching their limits," yet the industry remains optimistic due to Nvidia's past performance [3]. Group 2: New Chip Series - The specific models of the new chips have not been disclosed, but speculation suggests they may come from two major series: the Rubin series derivatives and the next-generation Feynman series [3]. - The Rubin series, which includes six new chip designs, has already been mass-produced, while the Feynman series is described as "revolutionary," exploring wide integration with SRAM and potential 3D stacking technology [3]. Group 3: AI Computing Needs - Nvidia is adapting to quarterly changes in AI computing demands, shifting focus from model pre-training with the Hopper and Blackwell series to inference scenarios with the Grace Blackwell Ultra and Vera Rubin series [3]. - The new products are expected to specifically address latency and memory bandwidth bottlenecks, which are critical for AI applications [3]. Group 4: Strategic Positioning - Huang emphasized that extensive collaboration and investment are key to Nvidia's continued leadership, as the company is positioning itself across the entire AI industry chain, including energy, semiconductors, and data centers [3].

Core Viewpoint - Tesla is strategically recruiting semiconductor talent in South Korea to strengthen its supply chain, accelerate production, and enhance its AI chip development capabilities [3][8]. Group 1: Recruitment Strategy - Tesla's recruitment announcement emphasizes the need for engineers in chip design, manufacturing, and AI software, highlighting the importance of solving challenging technical problems as a core evaluation criterion [1][6]. - The recruitment is part of a broader strategy to establish a team that can work closely with Samsung, facilitating a "co-located R&D" model to shorten the chip design cycle to an aggressive nine months [4][5]. Group 2: Supply Chain and Manufacturing - The decision to recruit in South Korea is driven by the physical proximity to key suppliers, particularly Samsung, with whom Tesla has signed a $16.5 billion long-term manufacturing agreement for AI chips [3][5]. - The AI5 chip, critical for Tesla's operations, will utilize Samsung's 3nm GAA process, ensuring that the production meets Tesla's urgent needs [3][4]. Group 3: Technological Advantages - South Korea is recognized as a global hub for HBM (High Bandwidth Memory) technology, essential for processing vast amounts of data required for Tesla's Full Self-Driving (FSD) capabilities [5][6]. - By establishing a local team, Tesla aims to optimize the interface performance between chips and memory, enhancing overall computational power and efficiency [5][6]. Group 4: Talent Pool and Geopolitical Considerations - South Korea boasts a top-tier semiconductor talent pool, driven by companies like Samsung and SK Hynix, which cover the entire semiconductor value chain [6][7]. - The recruitment strategy also considers geopolitical factors, as South Korea is a U.S. ally, providing a safer environment for technology transfer and intellectual property protection [7][8]. Group 5: Future Manufacturing Plans - The recruitment effort is aligned with Tesla's long-term goal of establishing its own trillion-dollar wafer fabrication plant (TeraFab), gradually reducing reliance on external suppliers [8].